NASP/cuda-samples
85
0
Fork 0
mirror of https://github.com/NVIDIA/cuda-samples.git synced 2026-10-26 06:38:16 +08:00
Code Issues Packages Projects Releases Wiki Activity

Apply consistent code formatting across the repo. Add clang-format and pre-commit hooks.

Browse Source
This commit is contained in:
Rob Armstrong
2025-03-27 10:30:07 -07:00
parent 2cd58fbc9a
commit ceab6e8bcc
782 changed files with 107230 additions and 106548 deletions
Download Patch File Download Diff File Expand all files Collapse all files

124
Samples/2_Concepts_and_Techniques/inlinePTX_nvrtc/inlinePTX.cpp
View File

@@ -30,8 +30,8 @@
*/
// System includes
#include <stdio.h>
#include <assert.h>
#include <stdio.h>
// CUDA runtime
#include <cuda_runtime.h>
@@ -40,63 +40,69 @@
// helper functions and utilities to work with CUDA
#include <helper_functions.h>
void sequence_cpu(int *h_ptr, int length) {
for (int elemID = 0; elemID < length; elemID++) {
h_ptr[elemID] = elemID % 32;
}
}
int main(int argc, char **argv) {
printf("CUDA inline PTX assembler sample\n");
char *cubin, *kernel_file;
size_t cubinSize;
kernel_file = sdkFindFilePath("inlinePTX_kernel.cu", argv[0]);
compileFileToCUBIN(kernel_file, argc, argv, &cubin, &cubinSize, 0);
CUmodule module = loadCUBIN(cubin, argc, argv);
CUfunction kernel_addr;
checkCudaErrors(cuModuleGetFunction(&kernel_addr, module, "sequence_gpu"));
const int N = 1000;
int *h_ptr = (int *)malloc(N * sizeof(int));
dim3 cudaBlockSize(256, 1, 1);
dim3 cudaGridSize((N + cudaBlockSize.x - 1) / cudaBlockSize.x, 1, 1);
CUdeviceptr d_ptr;
checkCudaErrors(cuMemAlloc(&d_ptr, N * sizeof(int)));
void *arr[] = {(void *)&d_ptr, (void *)&N};
checkCudaErrors(cuLaunchKernel(kernel_addr, cudaGridSize.x, cudaGridSize.y,
cudaGridSize.z, /* grid dim */
cudaBlockSize.x, cudaBlockSize.y,
cudaBlockSize.z, /* block dim */
0, 0, /* shared mem, stream */
&arr[0], /* arguments */
0));
checkCudaErrors(cuCtxSynchronize());
sequence_cpu(h_ptr, N);
int *h_d_ptr = (int *)malloc(N * sizeof(int));
checkCudaErrors(cuMemcpyDtoH(h_d_ptr, d_ptr, N * sizeof(int)));
bool bValid = true;
for (int i = 0; i < N && bValid; i++) {
if (h_ptr[i] != h_d_ptr[i]) {
bValid = false;
void sequence_cpu(int *h_ptr, int length)
{
for (int elemID = 0; elemID < length; elemID++) {
h_ptr[elemID] = elemID % 32;
}
}
printf("Test %s.\n", bValid ? "Successful" : "Failed");
checkCudaErrors(cuMemFree(d_ptr));
return bValid ? EXIT_SUCCESS : EXIT_FAILURE;
}
int main(int argc, char **argv)
{
printf("CUDA inline PTX assembler sample\n");
char *cubin, *kernel_file;
size_t cubinSize;
kernel_file = sdkFindFilePath("inlinePTX_kernel.cu", argv[0]);
compileFileToCUBIN(kernel_file, argc, argv, &cubin, &cubinSize, 0);
CUmodule module = loadCUBIN(cubin, argc, argv);
CUfunction kernel_addr;
checkCudaErrors(cuModuleGetFunction(&kernel_addr, module, "sequence_gpu"));
const int N = 1000;
int *h_ptr = (int *)malloc(N * sizeof(int));
dim3 cudaBlockSize(256, 1, 1);
dim3 cudaGridSize((N + cudaBlockSize.x - 1) / cudaBlockSize.x, 1, 1);
CUdeviceptr d_ptr;
checkCudaErrors(cuMemAlloc(&d_ptr, N * sizeof(int)));
void *arr[] = {(void *)&d_ptr, (void *)&N};
checkCudaErrors(cuLaunchKernel(kernel_addr,
cudaGridSize.x,
cudaGridSize.y,
cudaGridSize.z, /* grid dim */
cudaBlockSize.x,
cudaBlockSize.y,
cudaBlockSize.z, /* block dim */
0,
0, /* shared mem, stream */
&arr[0], /* arguments */
0));
checkCudaErrors(cuCtxSynchronize());
sequence_cpu(h_ptr, N);
int *h_d_ptr = (int *)malloc(N * sizeof(int));
checkCudaErrors(cuMemcpyDtoH(h_d_ptr, d_ptr, N * sizeof(int)));
bool bValid = true;
for (int i = 0; i < N && bValid; i++) {
if (h_ptr[i] != h_d_ptr[i]) {
bValid = false;
}
}
printf("Test %s.\n", bValid ? "Successful" : "Failed");
checkCudaErrors(cuMemFree(d_ptr));
return bValid ? EXIT_SUCCESS : EXIT_FAILURE;
}

17
Samples/2_Concepts_and_Techniques/inlinePTX_nvrtc/inlinePTX_kernel.cu
View File

@@ -25,15 +25,16 @@
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
extern "C" __global__ void sequence_gpu(int *d_ptr, int length) {
int elemID = blockIdx.x * blockDim.x + threadIdx.x;
extern "C" __global__ void sequence_gpu(int *d_ptr, int length)
{
int elemID = blockIdx.x * blockDim.x + threadIdx.x;
if (elemID < length) {
unsigned int laneid;
if (elemID < length) {
unsigned int laneid;
// This command gets the lane ID within the current warp
asm("mov.u32 %0, %%laneid;" : "=r"(laneid));
// This command gets the lane ID within the current warp
asm("mov.u32 %0, %%laneid;" : "=r"(laneid));
d_ptr[elemID] = laneid;
}
d_ptr[elemID] = laneid;
}
}